A Treebank-driven Creation of an OntoValence Verb Lexicon for Bulgarian
Petya Osenova, Kiril Simov, Laska Laskova, Stanislava Kancheva
Linguistic Modelling Department, IICT-BAS
Acad. G.Bonchev 25A, 1113 Sofia, Bulgaria
{petya, kivs, laska, stanislava}@bultreebank.org
Abstract
The paper presents a treebank-driven approach to the construction of a Bulgarian valence lexicon with ontological restrictions over the
inner participants of the event. First, the underlying ideas behind the Bulgarian Ontology-based lexicon are outlined. Then, the
extraction and manipulation of the valence frames is discussed with respect to the BulTreeBank annotation scheme and DOLCE
ontology. Also, the most frequent types of syntactic frames are specified as well as the most frequent types of ontological restrictions
over the verb arguments. The envisaged application of such a lexicon would be: in assigning ontological labels to syntactically parsed
corpora, and expanding the lexicon and lexical information in the Bulgarian Resource Grammar.
Keywords: Ontology-based Lexicon, Valence Lexicon, Treebank
1. Introduction
There exist various strategies towards constructing
valence dictionaries for a specific language. Some of these
resources are being created together with the extension of
a Treebank (Hinrichs and Telljohann, 2009) or comprise
the most frequent verbs, evaluated on a large corpus
(Zabokrtsky and Lopatkova, 2007). Also, valence
dictionaries keep closer to the surface syntactic level
(Hinrichs and Telljohann 2009) or elaborate on more
semantically
oriented
representations
(FrameNet,
PropBank, PDT-VALLEX, etc.).
The Prague strategy for building valence lexicons has
been adapted also to Arabic (Bielický and Smrž, 2008)
and Croatian (Agic et al, 2010) or for parallel lexicons,
such as the English-Czech valency lexicon attempt,
reported in (Šindlerová and Bojar, 2009). (Agic et al,
2010) claim that 1923 verb valency frames for 594
different lemmas have been extracted. Although the data
in Croatian is smaller, the ratio between the lemma and
the frames is comparable to our data, excluding in both
cases the valencies of the verb “to be”. Another approach
is taken for Danish – a combined representation of the
valency information is presented, which collapses the
LFG functions with HPSG categories (Amussen and
Ørsnes, 2005). An approach similar to ours with respect to
a Treebank-driven lexicon has been taken for Latin
(McGillivray and Passarotti, 2009). They follow the
notation of the Treebank itself, and thus the lexicon is
more data-dependent. In our case, the extracted frames
have been post-edited, if necessary, and also a new layer
of ontological abstraction over the arguments has been
added.
We aim at constructing a valency lexicon, which covers
the verbs in the syntactically analyzed corpus of Bulgarian
– BulTreeBank (www.bultreebank.org). Although the
surface syntactic structure is adopted, the lexicon consists
of ontological constraints. In this respect our strategy
resembles the approach in (Hinrichs and Telljohann,
2009). However, it also adds semantic information, but
rather based on a formal ontology than on thematic roles
or topic frames. Thus, our mechanism is more in lines
with the KYOTO Project (Vossen and Rigau, 2010) which
also uses DOLCE as a top ontology and OntoWordNet as
a mapping layer between the lexicon and the ontology.
This paper reports our observations on 3283 verb lemmas
in BulTreeBank. The number of distinct valence frames
for these lemmas is 6469. This means that the average is
almost 3 valence frames per lemma.
2. Bulgarian Ontology-based Lexicon
The valence lexicon presented in this paper is a part the
Bulgarian Ontology-based Lexicon (BOL) – (Simov and
Osenova, 2010). BOL is constructed on the basis of
ontology-to-text relation which interrelates an ontology, a
lexicon, an annotation grammar and text. The lexicon
exploits the relation between the ontology and the text.
We assume that the ontology represents the conceptual
meaning of the lexical units. The lexicon is mapped to the
ontology in order the lexical units to be connected to their
conceptual meaning. Additionally, the lexicon contains
linguistic
knowledge,
such
as
phonological,
morphological, and syntactic one. The mapping from the
lexicon to the ontology is done by two relations equality
and subsumption. The first relation is used when the
concept (relation) in the ontology corresponds exactly to
the conceptual meaning of the lexical unit. The second
relation is used when the meaning of the lexical unit is a
subconcept (subrelation) of the corresponding element in
the ontology. The usage of two relations allows us to
construct the ontology and the lexicon relatively
independent of each other in the sense that the coverage of
the lexicon can be extended more easily than the coverage
of the ontology.
The current version of BOL is based on DOLCE ontology
(Masolo et al, 2003) extended with concepts from
OntoWordNet (Gangemi et al, 2003) - a version 1.6 of
WordNet aligned to DOLCE. We selected DOLCE
Ontology as upper ontology for several reasons: (1) it is
constructed on rigorous basis which reflects the
OntoClean methodology (Guarino and Welty, 2002); (2) it
is represented in OWL-DL. We assume that the middle
layer of OntoWordNet contains concepts that are better
2636
understood by people. Thus the alignment between the
two ontologies facilitates the understanding of the
concepts in the upper ontology. The concepts from
OntoWordNet are additionally restricted by synsets
presented in EuroWordNet Base Concepts (BC) (Vossen, 1999) and Core WordNet 1 (CWN contains 5000
synsets from WordNet on the basis of analysis of British
National Corpus). We first used the intersection of both
selections of WordNet synsets. It contains 1504 synsets.
The corresponding concepts from OntoWordNet were
extracted and their alignment to DOLCE was used as a
first version of the ontology on which BOL has been
constructed. After we completed the Bulgarian lexicon for
the concepts from DOLCE and the intersection between
CWN and BC, we proceeded with rest of the concepts
from CWN. In addition, the lexicon was extended with
lexical units extracted from the Bulgarian National
Referent Corpus. These lexical units were ranked on the
basis of their frequency in the corpus and the number of
the documents in which they occurred. The last step
ensures that BOL will reflect in a better way the
Bulgarian-specific conceptualization. Thus, we assume
that an ontology, which consists of concepts and relations
from DOLCE and which is extended with concepts from
CWN, is abstract enough to provide a language
independent conceptual model for the construction of
ontology-based lexicons. We might envisage that most of
the lexical items in BOL would require more language
dependent concepts. Keeping a larger language
independent ontology might require more sophisticated
mapping relations between the lexicon and the ontology.
In (Simov and Osenova, 2010) we suggested that the
appropriate information for verb lexical units is
represented in two ways: (1) in the ontology each verb is
connected to an event concept related to the meaning of
the verb. In the ontology all the participants (irrespectivly
of whether they are considered to be arguments, adjuncts,
etc.) are represented as such via appropriate relations; (2)
the linguistic behavior is encoded in the lexicon as a set of
frames. These frames determine the role of each
participant in the given event. Unfortunately, there is no
available ontology with such detailed information on
events. Thus the presented work is in the direction of
constructing a verb lexicon according to the above
guidelines of a mapping between the lexicon and the
ontology. Later on, it will be used for the construction of
an event ontology.
To sum up, we consider ontologies and lexicons as
artifacts reflecting the human abilities for representing,
processing and managing linguistic and conceptual
knowledge. As such, they are not complete and
exhausting. This requires we to proceed in different ways
depending on the available information. In the next
sections we present how we use the information from the
syntactic treebank of Bulgarian - BulTreeBank in order to
construct a verbal lexicon in accordance with the model,
outlined above.
3. OntoValence Lexicon Extraction and
Manipulation
The valence lexicon has been acquired from the
syntactically
annotated
corpus
BulTreeBank.
BulTreeBank consists of 15 000 sentences (214 000
tokens). It includes predominantly newspaper texts, but
also prose and administrative documents. All the verbs
have been extracted together with the sentences they have
been used in. Then they have been lemmatized and sorted
by the lemma marker. The extracted valence frames
present the predefined decisions of the annotators, who
followed the BulTreeBank annotation scheme. Also, a
default valence frame has been inserted, which presents a
predicate with a SUBJ, DIROBJ and INDOBJ (NP
predicate N PP). In the BulTreeBank annotation it is
presented in the following way: (VPS NP (VPC V NP
PP))), where VPS stands for a head-subject phrase and
VPC stands for a head-complement phrase2. Thus, the
human expert received a list of lemmas, accompanied by
the respective sentence, which contains the annotated
frame, as well as the inserted default frame. The reasons
for choosing such an approach are as follows:
1. The pre-annotated frames in BulTreeBank might
differ syntactically from our present postulations
of constructing valence frames due to an error or
different view;
2. The pure copying of the annotated frame, which
might be considered a trivial step, has been
abandoned, since our aim is to add also
ontological constraints.
Thus, the default frame is viewed as a medium level to the
ontological abstraction.
Fig. 1: Original representation of a sentence tree in
BulTreeBank. The green rounded rectangle depicts the
lemma node of the verbal head of the sentence.
In Fig. 1 one of the usages of the verb назначавам
‘appoint - imperfective’ is presented in the BulTreeBank
visualization format. The gloss of the sentence is: Bluethe appoint officially area leader. The translation is: The
blue team ex officio appoints an area leader. In Fig. 2 the
2
1
http://wordnet.princeton.edu/wordnet/download/standoff/
Note that the canonical order is Subject – Verb – Direct
Object – Indirect Object.
2637
inserted default frame is presented, which says:
[SOMEBODY appoints SOMEONE for SOMETHING].
Then the expert can modify the frame according to the
adopted principles (more about the principles, see next
section). After modifying the frame (in this case –
deleting the PP), the expert adds the semantic constraints
over the arguments in accordance with the DOLCE
ontology. Such a modified and ontologically constrained
frame for the sentence, presented in Fig. 1, is showed in
Fig. 3. There the syntactic labels within the frame are
adapted to the scheme: SUBJ predicate DIROBJ, but the
lexical values of the arguments are replaced by
ontological concepts. Thus, the ontologically constrained
scheme is: ORGANIZATION [appoint - lemma]
PERSON.
importance is justified by the fact that 920 verb lemmas
out of the processed 2180 have occurred in BulTreeBank
only once; 313 lemmas have occurred 2 times; 200
lemmas – 3 times; 115 – 4 times and 94 – 5 times. Thus,
such cases constitute approximately 75 %, although the
most frequent verbs are well covered by the other 25 %
(for example, the verb гледам ‘look at’ has 93 frames,
extracted from BulTreeBank). We are aware of the fact,
that the real picture is much more complex, and for that
reason we envisage further investigations and elaborations
over the lexicon in the future.
4. OntoValence Lexicon Architecture and
Principles
Our approach to the valence lexicon follows the syntactic
annotation scheme in BulTreeBank. Here is the table with
the labels, which represent the main grammatical relations
between the predicate and the participant, and also
between other heads and their dependants:
Label
Description
VPA
head (verb)-adjunct
VPC
head(verb)-complement
VPS
head(verb)-subject
NPA
head(noun)-adjunct
NPC
head(noun)-complement
PP
head(preposition)-complement
PPA
head(preposition)-adjunct
APC
head(adjective)-complement
APA
head(adjective)-adjunct
AdvC
head(adverb)-complement
AdvA
head(adverb)-adjunct
Table 1: Description of the syntactic labels in
BulTreeBank
Fig. 2. The default frame included for each instance of a
verb. In this way some manual work is saved.
Fig. 3: The resulting frame
The extracted annotated frames from BulTreeBank are
18081. However, since our aim is to cover as many
syntactic and ontological variations per verb meaning as
possible, some additional example material has been
extracted also from the Bulgarian National Reference
Corpus. For the moment this step was executed only when
the examples per verb meaning are less than 5. Its
Apart from this, the syntactic architecture is more similar
to (Hinrichs and Telljohann, 2009) than to (Zabokrtsky
and Lopatkova, 2007). This means that we encode the
surface behavior of the verb occurrence and thus the
valence frame is kept to the surface syntax. In such a
framework the pro-drops of any kinds are also presented
within the frames. The frame considers the clausal
complements as well. We encode the verb usage in active
voice, while marking the passive occurrence in the
lemmatized wordform within the sentence. Contrary to
(Zabokrtsky and Lopatkova, 2007), the verbs in perfective
and imperfective aspect are considered separate lemmas.
As a rule, the frame includes only the inner participants
(semantically obligatory for the event or situation,
presented by the predicate, but might be unexpressed on
the surface level). Thus, the lexical meaning of the verb is
crucial for the frame creation. Our model follows the
ideas in (Pustejovsky, 1998: 63), i.e. we include the
TRUE
ARGUMENTS
(syntactically
realized),
DEFAULT ARGUMENTS (part of the semantics, but not
necessarily expressed) and SHADOW ARGUMENTS
(semantically incorporated into the lexical item), but not
TRUE ADJUNCTS (part of situational interpretation
2638
rather than of lexical one). Since the distinction between
complements and adjuncts is still problematic in the
theories of argument realization, for some unclear cases
an extended frame has been proposed. One problem is
whether some non-dative PPs, which seem to be
synonymic to the dative PP, are arguments, or not: говоря
на Иван (speak-I to Ivan ‘I speak to Ivan’) and говоря
пред публика (speak-I in front of audience ‘I speak to
some audience’). Thus, such extended frames have been
added for the verb гледам ‘look at’, among other verbs.
As is was already mentioned, our lexicon uses the
DOLCE ontology for constraining the inner participants.
This process is manual. The annotators have been
instructed to use the most abstract notion that keeps the
differentia specifica of the specific lexical meaning. In
cases when there is not such a notion, it has been added.
Thus, also the middle and domain layers of the top
ontology are extended. Our work is also facilitated by the
fact that parts of DOLCE have been integrated into
WordNet via OntoWordNet.
position. The most frequent type of clause that takes a
complement position is the so called ‘da-clause’, which in
general is considered an analytical counterpart of the Old
Bulgarian infinitive. The subject-null frame outnumbers
the subject explicit one.
N
1.
2.
Let us give some more specific information on the valence
frames of the processed lemmas. The most frequent
syntactic frames with the number of their occurrences are
presented in Table 2:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
Syntactic Frame Type
Number of
Frame
Occurrences
4089
3122
Predicate – Direct Object (NP)
Subject (NP) – Predicate –
Direct Object (NP)
Subject (NP) – Predicate
1339
Subject (NP) – Predicate –
1243
Indirect Object (PP)
Predicate
1082
Predicate – Direct Object (NP)
1013
– Indirect Object (PP)
Predicate – Indirect Object
888
(PP)
Predicate – Complement
807
(CLDA)
Subject (NP) - Predicate –
695
Direct Object (NP) – Indirect
Object (PP)
Subject (NP) - Predicate –
643
Complement (CLDA)
Table 2: Frequency of syntactic Frames
The most frequent frame is the one with a pro-drop SUBJ
and with an explicit Direct Object (NP). This fact proves
that the canonical Bulgarian utterance is a subject-null
one. The next most frequent frame is the one with a
realized SUBJ (NP) and a realized Direct Object (NP).
The intransitive frame takes the third position. It is worth
mentioning that also intransitive pro-drop frames seem to
occur often. In the frequency table they take the 5 th
Ontological Label
No Ontological
Restrictions
EVENT
Predicate – Complement
(CLDA)
3. Subject (NP) – Predicate
PERSON
4. Predicate – Direct Object
PERSON
(NP)
5. Subject (NP) - Predicate – PERSON - EVENT
Complement (CLDA)
6. Predicate – Direct Object
OBJECT
(NP)
7. Subject (NP) - Predicate – PERSON –
Direct Object (NP) –
ARTEFACT – (for)
Indirect Object (PP)
OBJECT
8. Subject (NP) – Predicate – PERSON Direct Object (NP)
PERSON
9. Predicate – Direct Object
SOCIAL OBJECT
(NP)
10. Subject (NP) – Predicate – PERSON Direct Object (NP)
OBJECT
Table 3: Frequency of ontological constraints
5. Some Observations over the current state
of the OntoValence Lexicon
N
Syntactic Frame
Predicate
As it can be seen in Table 3 above, from an ontological
point of view, the most frequent type is the type without
any restrictions. It is followed by a clausal complement,
which is constrained as an EVENT. Only then the
intransitive type with an explicit subject comes, with the
restriction PERSON. The same restriction holds for the
transitive subject-null frame, which is at the 4 th place.
Thus, the following restrictions over participants come in
a frequency order:
EVENT > PERSON > OBJECT > ARTEFACT >
SOCIAL OBJECT
It would be interesting to make also observations on the
frequency of participants within the sub-events within the
EVENT.
The 10 most frequent verbs with respect to the
occurrences in texts are as follows: the modal verb мога
‘can, be able to’; имам/има ‘have/there is’; нямам ‘do
not have’; кажа ‘say’; искам ‘want’; направя ‘doperfective’; правя ‘do-imperfecive’; знам ‘know’; заявя
‘declare, announce-perfective’; взема ‘take-perfective’;
видя ‘see-perfective’. Excluding the most frequent verb
съм ‘to be’, the next to come is another representative of
the stopword list, namely the modal verb мога ‘can’.
Closely to it are the verbs of possession and existence, and
the modal verb искам ‘want’. The following verbs show
some domain dependency in a corpus with predominantly
newspaper texts. This holds especially for the verbs say
and announce.
2639
6. Conclusions
The OntoValency lexicon has been fully processed with
respect to its coverage – both on syntactic and ontological
layers. However, more efforts are needed for testing the
correct level of abstraction for the ontological labeling,
and also the handling of the metaphorical usages of a
specific verb meaning.
The main problems in inter-annotator agreement seem to
be: the granularity of concept mappings as well as the
metaphorical and idiomatic usages. Concerning the
granularity, since the annotation has been example-based,
some annotators preferred to keep the most specific label,
while others provided some more abstract concept.
The metaphorical and idiomatic usages raised the question
of the adequate concept mapping. In both cases a mapping
is required also of whole phrases to concepts.
Presently, the OntoValency lexicon is envisaged to be
used for at least two tasks:
1. Annotation of syntactically parsed texts with
ontological constraints, and
2. Expanding the lexicon of the HPSG Bulgarian
Resource Grammar with valence information and
later on – incorporating the ontological
mappings.
The completion of the first task would support the various
applications of knowledge extraction, while the expanded
lexicon would improve massively the performance of the
HPSG grammar of Bulgarian.
A validation of the selected ontological restrictions over
the participants will be performed in two steps:
1. Mapping of the lexical units in BulTreeBank that
denote the participants in the concrete sentences
to the ontology of BOL;
2. Annotation of new sentences extracted from the
corpus via a clustering on the dependency
analyses of the sentences.
The first step will provide evidences whether the
conceptual restrictions on the event participants are
compatible with the concepts mapped to the lexical units
in BOL. The second step will provide evidences for
completeness of the frameset for each verb.
7. Acknowledgements
The work reported in this paper has been supported by the
EU project EuroMatrixPlus.
References
Željko Agic, Krešimir Šojat, Marko Tadic. 2010. An
Experiment in Verb Valency Frame Extraction from
Croatian Dependency Treebank. In: Proceedings of
the 32nd International Conference on Information
Technology Interfaces, Zagreb, SRCE University
Computer Centre, University of Zagreb, 2010, pp. 5560.
Jørg Asmussen and Bjarne Ørsnes. 2005. Valency
information for dictionaries and NLP lexicons:
Adapting valency frames from The Danish Dictionary
to an LFG lexicon. In: Ferenc Kiefer and Julia Pajzs
(eds.): Papers in Computational Lexicography.
Proceedings of the 8th Conference on Computational
Lexicography, COMPLEX 2005. Budapest: Research
Institute for Linguistics. Hungarian Academy of
Sciences, 2005, pp. 28–39.
Viktor Bielický and Otakar Smrž. 2008. Building the
Valency Lexicon of Arabic Verbs. In: Proceedings of
the Sixth International Conference on Language
Resources and Evaluation. ELRA
Aldo Gangemi, Roberto Navigli, and Paola Velardi.
(2003). The OntoWordNet Project: extension and
axiomatization of conceptual relations in WordNet.
Meersman R, et al. (eds.), Proceedings of ODBASE03
Conference, Springer, 2003.
Nicola Guarino and Welty, C. (2002). Evaluating
Ontological
Decisions
with
OntoClean.
Communications of the ACM, 45(2): 61-65.
Erhard Hinrichs and Heike Telljohann. 2009.
Constructing a Valence Lexicon for a Treebank of
German. In: Frank Van Eynde, Anette Frank,
Koenraad De Smedt and Gertjan van Noord (eds.)
Proceedings of the 7th International Workshop on
Treebanks and Linguistic Theories (TLT 7), LOT,
Groningen, pp. 41-52.
Claudio Masolo, Stefano Borgo,. Aldo Gangemi, Nicola
Guarino, Alessandro Oltramari. 2002. Ontology Library
(final). WonderWeb Deliverable D18, December 2003.
http://www.loa-cnr.it/Publications.html.
Barbara McGillivray and Marco Passarotti. 2009. The
Development of the Index Thomisticus Treebank
Valency Lexicon. In: Proceedings of the EACL 2009
Workshop on Language Technology and Resources
for Cultural Heritage, Social Sciences, Humanities,
and Education –LaTeCH – SHELT&R 2009, pages
43–50, Athens, Greece, 30 March 2009.
James Pustejovsky. 1998. The Generative Lexicon. The
MIT Press.
Kiril Simov and Petya Osenova. 2010. Constructing of an
Ontology-based Lexicon for Bulgarian. In
Proceedings of LREC 2010.
Jana Šindlerová and Ondrej Bojar. 2009. Towards
English-Czech Parallel Valency Lexicon via
Treebank Examples. In: Proceedings of the Eighth
International Workshop on Treebanks and Linguistic
Theories, 4-5 December 2009, Milan, Italy, pp. 185197.
Piek Vossen and G. Rigau. 2010. Division of semantic
labour in the Global WorNet Grid. In: Proceedings of
the 5th Global WordNet Conference, Mumbai, India.
Piek Vossen 1999. Editor. EuroWordNet General
Document. Version 3, Final, July 19, 1999.
http://www.hum.uva.nl/~ewn
Zdenek Zabokrtsky and Marketa Lopatkova. 2007.
Valency Information in VALLEX 2.0: Logical
Structure of the Lexicon. The Prague Bulletin of
Mathematical Linguistics No. 87, pp. 41-60, 2007.
2640